Outlier detection and high-dimensional data analysis are both highly active research areas, with ever-growing interest and demand in fields of Machine Learning and Data Science. In this study, we demonstrate an outlier detection via information centrality: a method using a graph centrality measure as a feature indicating a deviation of a test sample. First, we show a continuity result of the test effective conductance, of which the test information centrality consists.Then we present an application of the test information centrality, with an experiment on representation vectors of real ECG data generated by the Event Reconstruction JEPA, a self-supervised learning (SSL) method for a multivariate time series. Additionally, we will have an intervening section where we demonstrate that Event Reconstruction JEPA shows a performance on par with state-of-the-art SSL models with greatly reduced computational resources, made available by its two-fold hierarchical structure, with the first part constructing a summarized representation, a temporal representative of channels, and the second part taking the temporal representation as a univariate time series input.

발표시간: 17:00 ~ 18:00